Percy Liang
Assistant Professor of Computer Science and, by courtesy, of Statistics
Bio
Fields: machine learning, natural language processing.
Topics: unsupervised learning, structured prediction, statistical learning theory, grounded language acquisition, compositional semantics, program induction.
Learning semantics: Natural language allows us to express complex ideas using a few words, but the actual semantics are rarely directly observed. We therefore model the expressive semantics of language as programs whose execution produces observed data, and develop algorithms to learn these programs from indirect supervision.
Academic Appointments
-
Assistant Professor, Computer Science
-
Assistant Professor (By courtesy), Statistics
Professional Education
-
BS, MIT (2004)
-
MEng, MIT (2005)
-
PhD, UC Berkeley (2011)
2015-16 Courses
- Artificial Intelligence: Principles and Techniques
CS 221 (Aut) - Statistical Learning Theory
CS 229T, STATS 231 (Win) -
Independent Studies (15)
- Advanced Reading and Research
CS 499 (Aut, Win, Spr, Sum) - Advanced Reading and Research
CS 499P (Aut, Win, Spr, Sum) - Curricular Practical Training
CS 390A (Spr, Sum) - Curricular Practical Training
CS 390B (Aut, Win, Spr, Sum) - Curricular Practical Training
CS 390C (Aut, Win, Spr, Sum) - Independent Project
CS 399 (Aut, Spr) - Independent Project
CS 399P (Aut) - Independent Work
CS 199 (Aut, Win, Spr) - Independent Work
CS 199P (Aut, Win, Spr) - Part-Time CPT
CS 390S (Aut) - Part-Time CPT
CS 390T (Win) - Part-Time Curricular Practical Training
CS 390Q (Spr) - Part-time Curricular Practical Training
CS 390P (Win, Spr) - Senior Project
CS 191 (Aut, Win, Spr) - Writing Intensive Senior Project (WIM)
CS 191W (Aut, Spr)
- Advanced Reading and Research
-
Prior Year Courses
2014-15 Courses
- Artificial Intelligence: Principles and Techniques
CS 221 (Aut) - Statistical Learning Theory
CS 229T, STATS 231 (Win)
2013-14 Courses
- Artificial Intelligence: Principles and Techniques
CS 221 (Aut) - Statistical Learning Theory
CS 229T, STATS 231 (Win)
2012-13 Courses
- Artificial Intelligence: Principles and Techniques
CS 221 (Aut) - Statistical Learning Theory
CS 229T, STATS 231 (Win)
- Artificial Intelligence: Principles and Techniques
All Publications
-
Drug screening using a library of human induced pluripotent stem cell-derived cardiomyocytes reveals disease-specific patterns of cardiotoxicity.
Circulation
2013; 127 (16): 1677-1691
Abstract
Cardiotoxicity is a leading cause for drug attrition during pharmaceutical development and has resulted in numerous preventable patient deaths. Incidents of adverse cardiac drug reactions are more common in patients with preexisting heart disease than the general population. Here we generated a library of human induced pluripotent stem cell-derived cardiomyocytes (hiPSC-CMs) from patients with various hereditary cardiac disorders to model differences in cardiac drug toxicity susceptibility for patients of different genetic backgrounds.Action potential duration and drug-induced arrhythmia were measured at the single cell level in hiPSC-CMs derived from healthy subjects and patients with hereditary long QT syndrome, familial hypertrophic cardiomyopathy, and familial dilated cardiomyopathy. Disease phenotypes were verified in long QT syndrome, hypertrophic cardiomyopathy, and dilated cardiomyopathy hiPSC-CMs by immunostaining and single cell patch clamp. Human embryonic stem cell-derived cardiomyocytes (hESC-CMs) and the human ether-a-go-go-related gene expressing human embryonic kidney cells were used as controls. Single cell PCR confirmed expression of all cardiac ion channels in patient-specific hiPSC-CMs as well as hESC-CMs, but not in human embryonic kidney cells. Disease-specific hiPSC-CMs demonstrated increased susceptibility to known cardiotoxic drugs as measured by action potential duration and quantification of drug-induced arrhythmias such as early afterdepolarizations and delayed afterdepolarizations.We have recapitulated drug-induced cardiotoxicity profiles for healthy subjects, long QT syndrome, hypertrophic cardiomyopathy, and dilated cardiomyopathy patients at the single cell level for the first time. Our data indicate that healthy and diseased individuals exhibit different susceptibilities to cardiotoxic drugs and that use of disease-specific hiPSC-CMs may predict adverse drug responses more accurately than the standard human ether-a-go-go-related gene test or healthy control hiPSC-CM/hESC-CM screening assays.
View details for DOI 10.1161/CIRCULATIONAHA.113.001883
View details for PubMedID 23519760
-
Abnormal Calcium Handling Properties Underlie Familial Hypertrophic Cardiomyopathy Pathology in Patient-Specific Induced Pluripotent Stem Cells
CELL STEM CELL
2013; 12 (1): 101-113
Abstract
Familial hypertrophic cardiomyopathy (HCM) is a prevalent hereditary cardiac disorder linked to arrhythmia and sudden cardiac death. While the causes of HCM have been identified as genetic mutations in the cardiac sarcomere, the pathways by which sarcomeric mutations engender myocyte hypertrophy and electrophysiological abnormalities are not understood. To elucidate the mechanisms underlying HCM development, we generated patient-specific induced pluripotent stem cell cardiomyocytes (iPSC-CMs) from a ten-member family cohort carrying a hereditary HCM missense mutation (Arg663His) in the MYH7 gene. Diseased iPSC-CMs recapitulated numerous aspects of the HCM phenotype including cellular enlargement and contractile arrhythmia at the single-cell level. Calcium (Ca(2+)) imaging indicated dysregulation of Ca(2+) cycling and elevation in intracellular Ca(2+) ([Ca(2+)](i)) are central mechanisms for disease pathogenesis. Pharmacological restoration of Ca(2+) homeostasis prevented development of hypertrophy and electrophysiological irregularities. We anticipate that these findings will help elucidate the mechanisms underlying HCM development and identify novel therapies for the disease.
View details for DOI 10.1016/j.stem.2012.10.010
View details for Web of Science ID 000313839500014
- Feature noising for log-linear structured prediction. 2013
- A data driven approach for algebraic loop invariants. 2013
- Spectral experts for estimating mixtures of linear regressions. 2013
- Semantic parsing on Freebase from question-answer pairs. 2013
- Dropout training as adaptive regularization. 2013
- Video event understanding using natural language descriptions. 2013
-
Genome Editing of Human Embryonic Stem Cells and Induced Pluripotent Stem Cells With Zinc Finger Nucleases for Cellular Imaging
CIRCULATION RESEARCH
2012; 111 (12): 1494-?
Abstract
Molecular imaging has proven to be a vital tool in the characterization of stem cell behavior in vivo. However, the integration of reporter genes has typically relied on random integration, a method that is associated with unwanted insertional mutagenesis and positional effects on transgene expression.To address this barrier, we used genome editing with zinc finger nuclease (ZFN) technology to integrate reporter genes into a safe harbor gene locus (PPP1R12C, also known as AAVS1) in the genome of human embryonic stem cells and human induced pluripotent stem cells for molecular imaging.We used ZFN technology to integrate a construct containing monomeric red fluorescent protein, firefly luciferase, and herpes simplex virus thymidine kinase reporter genes driven by a constitutive ubiquitin promoter into a safe harbor locus for fluorescence imaging, bioluminescence imaging, and positron emission tomography imaging, respectively. High efficiency of ZFN-mediated targeted integration was achieved in both human embryonic stem cells and induced pluripotent stem cells. ZFN-edited cells maintained both pluripotency and long-term reporter gene expression. Functionally, we successfully tracked the survival of ZFN-edited human embryonic stem cells and their differentiated cardiomyocytes and endothelial cells in murine models, demonstrating the use of ZFN-edited cells for preclinical studies in regenerative medicine.Our study demonstrates a novel application of ZFN technology to the targeted genetic engineering of human pluripotent stem cells and their progeny for molecular imaging in vitro and in vivo.
View details for DOI 10.1161/CIRCRESAHA.112.274969
View details for Web of Science ID 000311994700042
View details for PubMedID 22967807
-
Induced Pluripotent Stem Cells as a Disease Modeling and Drug Screening Platform
JOURNAL OF CARDIOVASCULAR PHARMACOLOGY
2012; 60 (4): 408-416
Abstract
Induced pluripotent stem cells (iPSCs) hold great hopes for therapeutic application in various diseases. Although ongoing research is dedicated to achieving clinical translation of iPSCs, further understanding of the mechanisms that underlie complex pathogenic conditions is required. Compared with other classical models for studying diseases, iPSCs provide considerable advantages. A newly emerging application of iPSCs is in vitro disease modeling, which can significantly improve the never-ending search for new pharmacological cures. Here, we will discuss current efforts to create iPSC-dependent patient-specific disease models. Furthermore, we will review the use of iPSCs for development and testing of new therapeutic agents and the implications for high-throughput drug screening.
View details for DOI 10.1097/FJC.0b013e318247f642
View details for Web of Science ID 000309977900012
View details for PubMedID 22240913
-
Modeling Pathogenesis in Familial Hypertrophic Cardiomyopathy Using Patient-Specific Induced Pluripotent Stem Cells
LIPPINCOTT WILLIAMS & WILKINS. 2012
View details for Web of Science ID 000312506400056
- Identifiability and unmixing of latent parse trees. 2012
- Learning dependency-based compositional semantics. 2011
- Learning minimal abstractions. 2011
- Scaling up abstraction refinement via pruning. 2011
- A dynamic evaluation of static heap abstractions. 2010
- Learning programs: a hierarchical Bayesian approach. 2010
- A game-theoretic approach to generating spatial descriptions. 2010
- Type-based MCMC. 2010
- A simple domain-independent probabilistic approach to generation. 2010
- On the interaction between norm and dimensionality: multiple regimes in learning. 2010
- Learning from measurements in exponential families. 2009
- Learning semantic correspondences with less supervision. 2009
- Probabilistic grammars and hierarchical Dirichlet processes. The Oxford Handbook of Applied Bayesian Analysis 2009
- Online EM for unsupervised models. 2009
- Asymptotically optimal regularization in smooth parametric models. 2009
-
Optimal team size and monitoring in organizations
ACCOUNTING REVIEW
2008; 83 (3): 789-822
View details for Web of Science ID 000256277400008
- A probabilistic approach to language change. 2008
- Agreement-based learning. 2008
- Learning bilingual lexicons from monolingual corpora. 2008
- Structure compilation: trading structure for features. 2008
- Analyzing the errors of unsupervised learning. 2008
- An asymptotic analysis of generative, discriminative, and pseudolikelihood estimators. 2008
- A probabilistic approach to diachronic phonology. 2007
- A permutation-augmented sampler for Dirichlet process mixture models. 2007
- Structured Bayesian nonparametric models with variational inference (tutorial). 2007
- The infinite PCFG using hierarchical Dirichlet processes. 2007
- An end-to-end discriminative approach to machine translation. 2006
- Alignment by agreement. 2006
- Linear programming in bounded tree-width Markov networks. 2005
- A data structure for maintaining acyclicity in hypergraphs. Massachusetts Institute of Technology Technical Report 2005
- Efficient geometric algorithms for parsing in two dimensions. 2005
- Methods and experiments with bounded tree-width Markov networks. Massachusetts Institute of Technology Technical Report 2004
- How much of a hypertree can be captured by windmills? Massachusetts Institute of Technology Technical Report 2003
-
INTERFEROMETRIC STUDIES OF THE JOVIAN ATMOSPHERIC PROBE FIELD
AMER INST PHYSICS. 1980: 1093-1094
View details for Web of Science ID A1980KP44100161
-
Saponins and sapogenins. III. The sapogenins obtained from chlorogalum pomeridianum
JOURNAL OF THE AMERICAN CHEMICAL SOCIETY
1935; 57 (1): 525-527
View details for Web of Science ID 000188361300171